I am currently working on an item embedding task in recommendation system and I want to evaluate the performance of the new embedding algorithm with the old ones. I have read some papers about graph embedding and almost every paper mentioned a normal method to evaluate the embeddings which is link prediction. But none of these papers described exactly how you do it. So my question is how to evaluate the embeddings using link prediction?
The algorithm I am trying to apply is:
First a directed graph is built on user click sequences, each node in the graph represents an item, and if a user once clicked item A then clicked B, there should be two nodes A and B and an edge A-B with weight of 1. When another user clicked A then clicked B, the weight of edge A-B is added by 1.
Then a new sequence dataset is generated by random walking the graph, using the outbound weights as the teleport probabilities.
Finally SkipGram is performed on the new sequences to generate the node embeddings.
As many papers mentioned, I removed a certain proportion of the edges in the graph as the positive samples of test set(e.g. 0.25) and randomly generated some fake edges as the negative ones. So what's next? Should I simply generate fake edges for the real edges in the training set, concatenate the embeddings of the two nodes on each edge, and build a common classifier such as logistic regression and test it on the test set? Or should I calculate the AUC on test set with cosine similarity of the two nodes and a label of 0/1 indicating if the two nodes are really connected? Or should I calculate the AUC with the sigmoided dot product of the embeddings of two nodes and a label of 0/1 indicating if the two nodes are really connected, since this is how you compute the probability at last layer?
# these are example describing the three methods above
item_emb = np.random.random(400).reshape(100, 4) # assume we have 100 items and have embedded them into a 4-dimension vector space.
test_node = np.random.randint(0, 100, size=200).reshape(100, 2) # assume we have 100 pairs of nodes
test_label = np.random.randint(0, 2, size=100).reshape(100, 1) # assume this is the label indicating if the pair of nodes are really connected
def test_A():
# use logistic regression
train_node = ... # generate true and fake node pairs in a similar way
train_label = ... # generate true and fake node pairs in a similar way
train_feat = np.hstack(
item_emb[train_node[:, 0]],
item_emb[train_node[:, 1]]) # concatenate the embeddings
test_feat = np.hstack(
item_emb[test_node[:, 0]],
item_emb[test_node[:, 1]]) # concatenate the embeddings
lr = sklearn.linear_models.LogisticRegression().fit(train_feat, train_label)
auc = roc_auc_score(test_label, lr.predict_proba(test_feat)[:,1])
return auc
def test_B():
# use cosine similarity
emb1 = item_emb[test_node[:, 0]]
emb2 = item_emb[test_node[:, 1]]
cosine_sim = emb1 * emb2 / (np.linalg.norm(emb1, axis=1)*np.linalg.norm(emb2,axis=1)
auc = roc_auc_score(test_label, cosine_sim)
return auc
def test_C():
# use dot product
# here we extract the softmax weights and biases from the training network
softmax_weights = ... # same shape as item_emb
softmax_biases = ... # shape of (item_emb.shape[0], 1)
embedded_item = item_emb[test_node[:, 0]] # target item embedding
softmaxed_context = softmax_weights[test_node[:, 1]] + softmax_biases
dot_prod = np.sum(embeded_item * softmaxed_context, axis=1)
auc = roc_auc_score(test_label, dot_prod)
return auc
I have tried the three method in several tests, and they are not always telling the same thing. Some parameter combinations perform better with testA() and bad in other metrics, some the opposite..etc. Sadly there is no such a parameter combination that out performs others in all three metrics...The question is which metric should I use?
You should investigate some implementations:
StellarGraph: Link prediction with node2vec+Logistic regression
AmpliGraph: Link prediction with ComplEx
Briefly, one should sample edges (not nodes!) from the original graph, remove them, and learn embeddings on that truncated graph. Then an evaluation is performed on removed edges.
Also, there are two possible cases:
All possible edges between any pair of nodes are labeled. In this case evaluation metric is ROC AUC, when we learn a classifier to distinguish positives and negatives edges.
Only positive (real) edges are observed. We don't know if rest pairs are connected or not in the real world. Here we generate negative (fake) nodes for every positive one. The task is considered as Entity ranking with next evaluation metrics:
Rank
Mean Rank (MR)
Mean Reciprocal Rank (MRR)
Hits#N
An example can be found in the paper, sections 5.1-5.3.
Related
I have a similar open question here on Cross Validated (though not implementation focused, which I intend this question to be, so I think they are both valid).
I'm working on a project that uses sensors to monitor a persons GPS location. The coordinates will then be converted to a simple-grid representation. What I want to try and do is after recording a users routes, train a neural network to predict the next coordinates, i.e. take the example below where a user repeats only two routes over time, Home->A and Home->B.
I want to train an RNN/LSTM with sequences of varying lengths e.g. (14,3), (13,3), (12,3), (11,3), (10,3), (9,3), (8,3), (7,3), (6,3), (5,3), (4,3), (3,3), (2,3), (1,3) and then also predict with sequences of varying lengths e.g. for this example route if I called
route = [(14,3), (13,3), (12,3), (11,3), (10,3)] //pseudocode
pred = model.predict(route)
pred should give me (9,3) (or ideally even a longer prediction e.g. ((9,3), (8,3), (7,3), (6,3), (5,3), (4,3), (3,3), (2,3), (1,3))
How do I feed such training sequences to the init and forward operations identified below?
self.rnn = nn.RNN(input_size, hidden_dim, n_layers, batch_first=True)
out, hidden = self.rnn(x, hidden)
Also, should the entire route be a tensor or each set of coordinates within the route a tensor?
I'm not very experienced with RNNs, but I'll give it a try.
A few things to pay attention to before we start:
1. Your data is not normalized.
2. The output prediction you want (even after normalization) is not bounded to [-1, 1] range and therefore you cannot have tanh or ReLU activations acting on the output predictions.
To address your problem, I propose a recurrent net that given a current state (2D coordinate) predicts the next state (2D coordinates). Note that since this is a recurrent net, there is also a hidden state associated with each location. At first, the hidden state is zero, but as the net sees more steps, it updates its hidden state.
I propose a simple net to address your problem. It has a single RNN layer with 8 hidden states, and a fully connected layer on to to output the prediction.
class MyRnn(nn.Module):
def __init__(self, in_d=2, out_d=2, hidden_d=8, num_hidden=1):
super(MyRnn, self).__init__()
self.rnn = nn.RNN(input_size=in_d, hidden_size=hidden_d, num_layers=num_hidden)
self.fc = nn.Linear(hidden_d, out_d)
def forward(self, x, h0):
r, h = self.rnn(x, h0)
y = self.fc(r) # no activation on the output
return y, h
You can use your two sequences as training data, each sequence is a tensor of shape Tx1x2 where T is the sequence length, and each entry is two dimensional (x-y).
To predict (during training):
rnn = MyRnn()
pred, out_h = rnn(seq[:-1, ...], torch.zeros(1, 1, 8)) # given time t predict t+1
err = criterion(pred, seq[1:, ...]) # compare prediction to t+1
Once the model is trained, you can show it first k steps and continue to predict the next steps:
rnn.eval()
with torch.no_grad():
pred, h = rnn(s[:k,...], torch.zeros(1, 1, 8, dtype=torch.float))
# pred[-1, ...] is the predicted next step
prev = pred[-1:, ...]
for j in range(k+1, s.shape[0]):
pred, h = rnn(prev, h) # note how we keep track of the hidden state of the model. it is no longer init to zero.
prev = pred
I put everything together in a colab notebook so you can play with it.
For simplicity, I ignored the data normalization here, but you can find it in the colab notebook.
What's next?
These types of predictions are prone to error accumulation. This should be addressed during training, by shifting the inputs from the ground truth "clean" sequences to the actual predicted sequences, so the model will be able to compensate for its errors.
I implemented a cosine-theta function, which calculates the relation between two articles. If two articles are very similar then the words should contain quite some overlap. However, a cosine theta score of 0.54 does not mean "related" or "not related". I should end up with a definitive answer which is either 0 for 'not related' or 1 for 'related'.
I know that there are sigmoid and softmax functions, yet I should find the optimal parameters to give to such functions and I do not know if these functions are satisfactory solutions. I was thinking that I have the cosine theta score, I can calculate the percentage of overlap between two sentences two (e.g. the amount of overlapping words divided by the amount of words in the article) and maybe some more interesting things. Then with the data, I could maybe write a function (what type of function I do not know and is part of the question!), after which I can minimize the error via the SciPy library. This means that I should do some sort of supervised learning, and I am willing to label article pairs with labels (0/1) in order to train a network. Is this worth the effort?
# Count words of two strings.
v1, v2 = self.word_count(s1), self.word_count(s2)
# Calculate the intersection of the words in both strings.
v3 = set(v1.keys()) & set(v2.keys())
# Calculate some sort of ratio between the overlap and the
# article length (since 1 overlapping word on 2 words is more important
# then 4 overlapping words on articles of 492 words).
p = min(len(v1), len(v2)) / len(v3)
numerator = sum([v1[w] * v2[w] for w in v3])
w1 = sum([v1[w]**2 for w in v1.keys()])
w2 = sum([v2[w]**2 for w in v2.keys()])
denominator = math.sqrt(w1) * math.sqrt(w2)
# Calculate the cosine similarity
if not denominator:
return 0.0
else:
return (float(numerator) / denominator)
As said, I would like to use variables such as p, and the cosine theta score in order to produce an accurate discrete binary label, either 0 or 1.
As said, I would like to use variables such as p, and the cosine theta score in order to produce an accurate discrete binary label, either 0 or 1.
Here it really comes down to what you mean by accuracy. It is up to you to choose how the overlap affects whether or not two strings are "matching" unless you have a labelled data set. If you have a labelled data set (I.e., a set of pairs of strings along with a 0 or 1 label), then you can train a binary classification algorithm and try to optimise based on that. I would recommend something like a neural net or SVM due to the potentially high dimensional, categorical nature of your problem.
Even the optimisation, however, is a subjective measure. For example, in theory let's pretend you have a model which out of 100 samples only predicts 1 answer (Giving 99 unknowns). Technically if that one answer is correct, that is a model with 100% accuracy, but which has a very low recall. Generally in machine learning you will find a trade off between recall and accuracy.
Some people like to go for certain metrics which combine the two (The most famous of which is the F1 score), but honestly it depends on the application. If I have a marketing campaign with a fixed budget, then I care more about accuracy - I would only want to target consumers who are likely to buy my product. If however, we are looking to test for a deadly disease or markers for bank fraud, then it's feasible for that test to be accurate only 10% of the time - if its recall of true positives is somewhere close to 100%.
Finally, if you have no labelled data, then your best bet is just to define some cut off value which you believe indicates a good match. This is would then be more analogous to a binary clustering problem, and you could use some more abstract measure such as distance to a centroid to test which cluster (Either the "related" or "unrelated" cluster) the point belongs to. Note however that here your features feel like they would be incredibly hard to define.
I'm building Kmeans in pytorch using gradient descent on centroid locations, instead of expectation-maximisation. Loss is the sum of square distances of each point to its nearest centroid. To identify which centroid is nearest to each point, I use argmin, which is not differentiable everywhere. However, pytorch is still able to backprop and update weights (centroid locations), giving similar performance to sklearn kmeans on the data.
Any ideas how this is working, or how I can figure this out within pytorch? Discussion on pytorch github suggests argmax is not differentiable: https://github.com/pytorch/pytorch/issues/1339.
Example code below (on random pts):
import numpy as np
import torch
num_pts, batch_size, n_dims, num_clusters, lr = 1000, 100, 200, 20, 1e-5
# generate random points
vector = torch.from_numpy(np.random.rand(num_pts, n_dims)).float()
# randomly pick starting centroids
idx = np.random.choice(num_pts, size=num_clusters)
kmean_centroids = vector[idx][:,None,:] # [num_clusters,1,n_dims]
kmean_centroids = torch.tensor(kmean_centroids, requires_grad=True)
for t in range(4001):
# get batch
idx = np.random.choice(num_pts, size=batch_size)
vector_batch = vector[idx]
distances = vector_batch - kmean_centroids # [num_clusters, #pts, #dims]
distances = torch.sum(distances**2, dim=2) # [num_clusters, #pts]
# argmin
membership = torch.min(distances, 0)[1] # [#pts]
# cluster distances
cluster_loss = 0
for i in range(num_clusters):
subset = torch.transpose(distances,0,1)[membership==i]
if len(subset)!=0: # to prevent NaN
cluster_loss += torch.sum(subset[:,i])
cluster_loss.backward()
print(cluster_loss.item())
with torch.no_grad():
kmean_centroids -= lr * kmean_centroids.grad
kmean_centroids.grad.zero_()
As alvas noted in the comments, argmax is not differentiable. However, once you compute it and assign each datapoint to a cluster, the derivative of loss with respect to the location of these clusters is well-defined. This is what your algorithm does.
Why does it work? If you had only one cluster (so that the argmax operation didn't matter), your loss function would be quadratic, with minimum at the mean of the data points. Now with multiple clusters, you can see that your loss function is piecewise (in higher dimensions think volumewise) quadratic - for any set of centroids [C1, C2, C3, ...] each data point is assigned to some centroid CN and the loss is locally quadratic. The extent of this locality is given by all alternative centroids [C1', C2', C3', ...] for which the assignment coming from argmax remains the same; within this region the argmax can be treated as a constant, rather than a function and thus the derivative of loss is well-defined.
Now, in reality, it's unlikely you can treat argmax as constant, but you can still treat the naive "argmax-is-a-constant" derivative as pointing approximately towards a minimum, because the majority of data points are likely to indeed belong to the same cluster between iterations. And once you get close enough to a local minimum such that the points no longer change their assignments, the process can converge to a minimum.
Another, more theoretical way to look at it is that you're doing an approximation of expectation maximization. Normally, you would have the "compute assignments" step, which is mirrored by argmax, and the "minimize" step which boils down to finding the minimizing cluster centers given the current assignments. The minimum is given by d(loss)/d([C1, C2, ...]) == 0, which for a quadratic loss is given analytically by the means of data points within each cluster. In your implementation, you're solving the same equation but with a gradient descent step. In fact, if you used a 2nd order (Newton) update scheme instead of 1st order gradient descent, you would be implicitly reproducing exactly the baseline EM scheme.
Imagine this:
t = torch.tensor([-0.0627, 0.1373, 0.0616, -1.7994, 0.8853,
-0.0656, 1.0034, 0.6974, -0.2919, -0.0456])
torch.argmax(t).item() # outputs 6
We increase t[0] for some, δ close to 0, will this update the argmax? It will not, so we are dealing with 0 gradients, all the time. Just ignore this layer, or assume it is frozen.
The same is for argmin, or any other function where the dependent variable is in discrete steps.
I'm using the ScikitLearn flavour of the DecisionTree.jl package to create a random forest model for a binary classification problem of one of the RDatasets data sets (see bottom of the DecisionTree.jl home page for what I mean by ScikitLearn flavour). I'm also using the MLBase package for model evaluation.
I have built a random forest model of my data and would like to create a ROC Curve for this model. Reading the documentation available, I do understand what a ROC curve is in theory. I just can't figure out how to create one for a specific model.
From the Wikipedia page the last part of the first sentence that I have marked in bold italics below is the one that is causing my confusion: "In statistics, a receiver operating characteristic (ROC), or ROC curve, is a graphical plot that illustrates the performance of a binary classifier system as its discrimination threshold is varied." There is more on the threshold value throughout the article but this still confuses me for binary classification problems. What is the threshold value and how do I vary it?
Also, in the MLBase documentation on ROC Curves it says "Compute an ROC instance or an ROC curve (a vector of ROC instances), based on given scores and a threshold thres." But doesn't mention this threshold anywhere else really.
Example code for my project is given below. Basically, I want to create a ROC curve for the random forest but I'm not sure how to or if it's even appropriate.
using DecisionTree
using RDatasets
using MLBase
quakes_data = dataset("datasets", "quakes");
# Add in a binary column as feature column for classification
quakes_data[:MagGT5] = convert(Array{Int32,1}, quakes_data[:Mag] .> 5.0)
# Getting features and labels where label = 1 is mag > 1 and label = 2 is mag <= 5
features = convert(Array, quakes_data[:, [1:3;5]]);
labels = convert(Array, quakes_data[:, 6]);
labels[labels.==0] = 2
# Create a random forest model with the tuning parameters I want
r_f_model = RandomForestClassifier(nsubfeatures = 3, ntrees = 50, partialsampling=0.7, maxdepth = 4)
# Train the model in-place on the dataset (there isn't a fit function without the in-place functionality)
DecisionTree.fit!(r_f_model, features, labels)
# Apply the trained model to the test features data set (here I haven't partitioned into training and test)
r_f_prediction = convert(Array{Int64,1}, DecisionTree.predict(r_f_model, features))
# Applying the model to the training set and looking at model stats
TrainingROC = roc(labels, r_f_prediction) #getting the stats around the model applied to the train set
# p::T # positive in ground-truth
# n::T # negative in ground-truth
# tp::T # correct positive prediction
# tn::T # correct negative prediction
# fp::T # (incorrect) positive prediction when ground-truth is negative
# fn::T # (incorrect) negative prediction when ground-truth is positive
I also read this question and didn't find it helpful really.
The task in binary classification is to give a 0/1 (or true/false, red/blue) label to a new, unlabeled, data-point. Most classification algorithms are designed to output a continuous real value. This value is optimized to be higher for points with known or predicted label 1, and lower for points with known or predicted label 0. To use this value to generate a 0/1 prediction, an additional threshold is used. Points with a value higher than threshold are predicted to be labeled 1 (and for lower than threshold a 0 label is predicted ).
Why is this setup useful? Because, sometimes mispredicting a 0 instead of a 1 is more costly, and then you can set the threshold low, making the algorithm output predict 1s more often.
In an extreme case when predicting 0 instead of a 1 costs nothing for the application, you can set the threshold at infinity, making it always output 0 (which is obviously the best solution, since it incurs no cost).
The threshold trick cannot eliminate errors from the classifier - no classifier in real-world problems is perfect or free from noise. What it can do is change the ratio between the 0-when-really-1 errors and 1-when-really-0 errors for the final classification.
As you increase the threshold, more points are classified with a 0 label. Consider a chart with the fraction of points classified with 0 on the x-axis, and the fraction of points with a 0-when-really-1 error on the y-axis. For each value of the threshold, plot a point for the resulting classifier on this chart. Plotting a point for all thresholds you get a curve. This is (some variant of) the ROC curve, which summarizes the abilities of the classifier. An often used metric for quality of classification is the AUC or area-under-curve of this chart, but in fact, the whole curve can be of interest in applications.
A summary like this appears in many texts on machine learning, which are a google query away.
Hope this clarifies the role of the threshold and its relation to ROC curves.
I am trying to use RBFNN for point cloud to surface reconstruction but I couldn't understand what would be my feature vectors in RBFNN.
Can any one please help me to understand this one.
A goal to get to this:
From inputs like this:
An RBF network essentially involves fitting data with a linear combination of functions that obey a set of core properties -- chief among these is radial symmetry. The parameters of each of these functions is learned by incremental adjustment based on errors generated through repeated presentation of inputs.
If I understand (it's been a very long time since I used one of these networks), your question pertains to preprocessing of the data in the point cloud. I believe that each of the points in your point cloud should serve as one input. If I understand properly, the features are your three dimensions, and as such each point can already be considered a "feature vector."
You have other choices that remain, namely the number of radial basis neurons in your hidden layer, and the radial basis functions to use (a Gaussian is a popular first choice). The training of the network and the surface reconstruction can be done in a number of ways but I believe this is beyond the scope of the question.
I don't know if it will help, but here's a simple python implementation of an RBF network performing function approximation, with one-dimensional inputs:
import numpy as np
import matplotlib.pyplot as plt
def fit_me(x):
return (x-2) * (2*x+1) / (1+x**2)
def rbf(x, mu, sigma=1.5):
return np.exp( -(x-mu)**2 / (2*sigma**2));
# Core parameters including number of training
# and testing points, minimum and maximum x values
# for training and testing points, and the number
# of rbf (hidden) nodes to use
num_points = 100 # number of inputs (each 1D)
num_rbfs = 20.0 # number of centers
x_min = -5
x_max = 10
# Training data, evenly spaced points
x_train = np.linspace(x_min, x_max, num_points)
y_train = fit_me(x_train)
# Testing data, more evenly spaced points
x_test = np.linspace(x_min, x_max, num_points*3)
y_test = fit_me(x_test)
# Centers of each of the rbf nodes
centers = np.linspace(-5, 10, num_rbfs)
# Everything is in place to train the network
# and attempt to approximate the function 'fit_me'.
# Start by creating a matrix G in which each row
# corresponds to an x value within the domain and each
# column i contains the values of rbf_i(x).
center_cols, x_rows = np.meshgrid(centers, x_train)
G = rbf(center_cols, x_rows)
plt.plot(G)
plt.title('Radial Basis Functions')
plt.show()
# Simple training in this case: use pseudoinverse to get weights
weights = np.dot(np.linalg.pinv(G), y_train)
# To test, create meshgrid for test points
center_cols, x_rows = np.meshgrid(centers, x_test)
G_test = rbf(center_cols, x_rows)
# apply weights to G_test
y_predict = np.dot(G_test, weights)
plt.plot(y_predict)
plt.title('Predicted function')
plt.show()
error = y_predict - y_test
plt.plot(error)
plt.title('Function approximation error')
plt.show()
First, you can explore the way in which inputs are provided to the network and how the RBF nodes are used. This should extend to 2D inputs in a straightforward way, though training may get a bit more involved.
To do proper surface reconstruction you'll likely need a representation of the surface that is altogether different than the representation of the function that's learned here. Not sure how to take this last step.